One-Shot Architecture Search and Transformation for Robust DOA Estimation

Abstract

Given the challenges of direction-of-arrival (DOA) estimation methods under low signal-to-noise ratios (SNRs), we propose a one-shot architecture search and transformation DOA estimation (OAST-DOA) framework for robust DOA estimation. First, by formulating the DOA estimation problem as a multilabel classification task, the multichannel training data are constructed from the real covariance matrix under low SNRs. A long short-term memory network is introduced as a controller to guide the process of architecture search and optimal cell selection. In addition, to reduce the computational complexity without compromising performance, the computationally intensive operations are transformed into more efficient alternatives within the optimal cell via architecture transformation. Simulation results show that the proposed OAST-DOA method has significant advantages for scenarios with low SNRs and a relatively small number of snapshots, and exhibits robustness against array model errors.